arm_gemm: SME: Remove artificial single-thread constraint on quantized int8 kernels.
Change-Id: I81b71ecc0d2e776d132091e074798a79b3141bec
Signed-off-by: David Mansell <David.Mansell@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/11085
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp
index c725815..7c860a2 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2019-2020, 2022-2023 Arm Limited.
+ * Copyright (c) 2019-2020, 2022-2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -80,7 +80,7 @@
{
GemmMethod::GEMM_INTERLEAVED,
"sme2_interleaved_nomerge_s8q_mopa_1VLx4VL",
- [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sme2() && args._maxthreads == 1 && ((qp.per_channel_requant && (qp.per_channel_left_shifts == nullptr)) || (!qp.per_channel_requant && (qp.per_layer_left_shift == 0)));},
+ [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sme2() && ((qp.per_channel_requant && (qp.per_channel_left_shifts == nullptr)) || (!qp.per_channel_requant && (qp.per_layer_left_shift == 0)));},
[](const GemmArgs &args, const Requantize32 &) { const auto VL = sme::get_vector_length<int32_t>();
return args._Msize <= VL || (2*VL < args._Msize && args._Msize <= 3*VL); },
[](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedPretransposedNoMergeQuantizedInline<cls_sme2_interleaved_nomerge_s8q_mopa_1VLx4VL, int8_t, int8_t>(args, qp); }
@@ -88,7 +88,7 @@
{
GemmMethod::GEMM_INTERLEAVED,
"sme2_interleaved_nomerge_s8q_mopa_4VLx1VL",
- [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sme2() && args._maxthreads == 1 && ((qp.per_channel_requant && (qp.per_channel_left_shifts == nullptr)) || (!qp.per_channel_requant && (qp.per_layer_left_shift == 0)));},
+ [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sme2() && ((qp.per_channel_requant && (qp.per_channel_left_shifts == nullptr)) || (!qp.per_channel_requant && (qp.per_layer_left_shift == 0)));},
[](const GemmArgs &args, const Requantize32 &) { const auto VL = sme::get_vector_length<int32_t>();
return args._Nsize <= VL || (2*VL < args._Nsize && args._Nsize <= 3*VL); },
[](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedPretransposedNoMergeQuantizedInline<cls_sme2_interleaved_nomerge_s8q_mopa_4VLx1VL, int8_t, int8_t>(args, qp); }
@@ -96,7 +96,7 @@
{
GemmMethod::GEMM_INTERLEAVED,
"sme2_interleaved_nomerge_s8q_mopa_2VLx2VL",
- [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sme2() && args._maxthreads == 1 && ((qp.per_channel_requant && (qp.per_channel_left_shifts == nullptr)) || (!qp.per_channel_requant && (qp.per_layer_left_shift == 0)));},
+ [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sme2() && ((qp.per_channel_requant && (qp.per_channel_left_shifts == nullptr)) || (!qp.per_channel_requant && (qp.per_layer_left_shift == 0)));},
nullptr,
[](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedPretransposedNoMergeQuantizedInline<cls_sme2_interleaved_nomerge_s8q_mopa_2VLx2VL, int8_t, int8_t>(args, qp); }
},
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp
index 6254ec6..3baf985 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2019-2020, 2022-2023 Arm Limited.
+ * Copyright (c) 2019-2020, 2022-2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -76,7 +76,7 @@
{
GemmMethod::GEMM_INTERLEAVED,
"sme2_interleaved_nomerge_u8q_mopa_1VLx4VL",
- [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sme2() && args._maxthreads == 1 && ((qp.per_channel_requant && (qp.per_channel_left_shifts == nullptr)) || (!qp.per_channel_requant && (qp.per_layer_left_shift == 0)));},
+ [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sme2() && ((qp.per_channel_requant && (qp.per_channel_left_shifts == nullptr)) || (!qp.per_channel_requant && (qp.per_layer_left_shift == 0)));},
[](const GemmArgs &args, const Requantize32 &) { const auto VL = sme::get_vector_length<uint32_t>();
return args._Msize <= VL || (2*VL < args._Msize && args._Msize <= 3*VL); },
[](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedPretransposedNoMergeQuantizedInline<cls_sme2_interleaved_nomerge_u8q_mopa_1VLx4VL, uint8_t, uint8_t>(args, qp); }
@@ -84,7 +84,7 @@
{
GemmMethod::GEMM_INTERLEAVED,
"sme2_interleaved_nomerge_u8q_mopa_4VLx1VL",
- [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sme2() && args._maxthreads == 1 && ((qp.per_channel_requant && (qp.per_channel_left_shifts == nullptr)) || (!qp.per_channel_requant && (qp.per_layer_left_shift == 0)));},
+ [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sme2() && ((qp.per_channel_requant && (qp.per_channel_left_shifts == nullptr)) || (!qp.per_channel_requant && (qp.per_layer_left_shift == 0)));},
[](const GemmArgs &args, const Requantize32 &) { const auto VL = sme::get_vector_length<int32_t>();
return args._Nsize <= VL || (2*VL < args._Nsize && args._Nsize <= 3*VL); },
[](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedPretransposedNoMergeQuantizedInline<cls_sme2_interleaved_nomerge_u8q_mopa_4VLx1VL, uint8_t, uint8_t>(args, qp); }
@@ -92,7 +92,7 @@
{
GemmMethod::GEMM_INTERLEAVED,
"sme2_interleaved_nomerge_u8q_mopa_2VLx2VL",
- [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sme2() && args._maxthreads == 1 && ((qp.per_channel_requant && (qp.per_channel_left_shifts == nullptr)) || (!qp.per_channel_requant && (qp.per_layer_left_shift == 0)));},
+ [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sme2() && ((qp.per_channel_requant && (qp.per_channel_left_shifts == nullptr)) || (!qp.per_channel_requant && (qp.per_layer_left_shift == 0)));},
nullptr,
[](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedPretransposedNoMergeQuantizedInline<cls_sme2_interleaved_nomerge_u8q_mopa_2VLx2VL, uint8_t, uint8_t>(args, qp); }
},